NNET Procedure
MITIGATEBIAS Statement
MITIGATEBIAS SENSITIVEVAR=variable PREDICTEDVARS=(variable-names) PREDICTEDEVENTS="event-list" <options>;
The MITIGATEBIAS statement enables PROC NNET to iteratively train a model while minimizing the bias metric that you specify. To use this statement, you must also specify the SAVESTATE statement and a categorical response (target) variable. You cannot also specify the AUTOTUNE statement or the CROSSVALIDATION statement.
You must specify the following options:
-
PREDICTEDEVENTS="event-list"
PEVENTS="event-list" specifies the events that correspond to each variable in the PREDICTEDVARS= option. You must specify the PREDICTEDEVENTS= and PREDICTEDVARS= options with the same number of items in the same order. Items in the event-list must be separated by the delimiter that you specify in the DELIMETER= option.
-
PREDICTEDVARS=(variable-names)
PVARS=(variable-names) specifies the names of the variables that contain the posterior probability for each level in model prediction that corresponds to the response (target) variable. You must specify the PREDICTEDVARS= and PREDICTEDEVENTS= options with the same number of items in the same order. Duplicate variables are not allowed.
- SENSITIVEVAR=variable
specifies the sensitive variable to use in order to reduce the value of the bias measurement that you specify in the BIASMETRIC= option.
You can also specify the following options:
- BIASMETRIC=bias-metric
-
specifies the type of bias measurement. You can specify the following bias-metrics:
- DEMOGRAPHICPARITY
specifies demographic parity as the bias measurement.
- EQUALIZEDODDS
specifies equalized odds as the bias measurement.
- EQUALOPPORTUNITY
specifies equal opportunity as the bias measurement.
- PREDICTIVEPARITY
specifies predictive parity as the bias measurement.
By default, BIASMETRIC=DEMOGRAPHICPARITY.
- BOUND=number
-
specifies the bound value for the exponentiated gradient reduction algorithm.
By default, BOUND=100.
-
DELIMITER="character"
DLM="character" -
specifies the delimiter to be used to separate events that you specify in the PREDICTEDEVENTS= option. A delimiter is used when event names contain embedded spaces (such as "Fuel Oil") or special characters (such as ";" or "*"). You must use quotation marks around character. You can specify only one delimiter; combinations of delimiters are not supported. The valid values of the DELIMITER= option are " " (space), ";" (semicolon), "*" (asterisk), "." (period), and "," (comma).
By default, DELIMITER=" " (space).
If you specify a delimiter other than " " (space), PROC NNET treats any spaces (including empty spaces) as valid characters. You are responsible for mapping events correctly in order to obtain correct results.
- LEARNINGRATE=number
-
specifies the step size to use in updating the exponentiated gradient reduction algorithm.
By default, LEARNINGRATE=0.01.
- LOGLEVEL=0 | 1 | 2
-
specifies the level of log information to print.
- 0
prints only warnings and errors.
- 1
prints some notes.
- 2
prints notes for each iteration.
By default, LOGLEVEL=1.
- MAXITER=number
-
specifies the maximum number of iterations to run the exponentiated gradient reduction algorithm.
By default, MAXITER=10.
- SEED=number
-
specifies the seed for the pseudorandom number generator. By default, the random number stream is based on the time of day from the computer’s clock. A 0 or negative seed value also results in random number streams that are based on the time of day from the computer’s clock. If you want a reproducible random number sequence between runs, specify a value greater than 0.
By default, SEED=0.
- TARGETEVENT="event"
-
specifies the formatted value of the response (target) variable that represents the event of interest.
By default, event is the first level that you specify in the PREDICTEDEVENTS= option.
- TOLERANCE=number
-
specifies the parity constraint violation tolerance. If you set number to 0, PROC NNET trains for the maximum number of iterations, which you specify in the MAXITER= option.
By default, TOLERANCE=0.005.
- TUNEBOUND
specifies that the bound value must be tuned.